Harness Engineering
Source-grounded pack with Agent Runtime Architecture Map, Context and State Contract, and Agent Run Evaluation Harness, plus 1 more.
Get this packGrounded in the complete Harness Engineering series.
Open the pack, choose a workflow, and follow its setup and input instructions. You can also open each workflow below.
10 source essays · 4 usable skills · reviewed Jul 2026
Use this pack when
Choose this pack when the work touches Harness Engineering and you need a bounded workflow with explicit inputs, a defined output, and human review gates.
Choose the job in front of you.
4 source-grounded skills. Load one workflow at a time.
Agent Runtime Architecture Map
Use this workflow when an agent workflow must move from a good demo or chat prompt into a reproducible runtime with explicit owners, stages, orchestration, context, tools, state, reviews, permissions, logs, failure handling, and learning loops.
Output specification
Return an agent runtime architecture with run_spec, stage_map, orchestration_contract, control_links, trace_and_replay, middleware_boundary, and operating_model.
Context and State Contract
Use this workflow when an agent workflow needs an explicit contract for authoritative context, exclusions, provenance, freshness, retrieval, compression, memory, state transitions, checkpoints, invalidation, and safe resumption across long or interrupted runs.
Output specification
Return a context and state contract with context_inventory, precedence_and_exclusions, freshness_conflict_rules, loading_compression, state_model, checkpoint_schema, and resumption_tests.
Agent Run Evaluation Harness
Use this workflow when an agent workflow needs replayable evaluation of context, decisions, tool calls, state, approvals, outcomes, safety, cost, latency, regressions, and operator learning rather than prompt-quality scores alone.
Output specification
Return an agent evaluation harness specification with evaluation_contract, trace_schema, case_suite, scoring_rubric, release_thresholds, analysis_views, and learning_loop.
Agent Runtime Safety Review
Use this workflow when an agent can call tools, access data, create side effects, or escalate decisions and needs recoverable tool interfaces, least privilege, sandboxes, proportional human review, failure boundaries, readable errors, stop controls, and incident response.
Output specification
Return an agent runtime safety review with action_inventory, tool_contracts, permission_matrix, failure_analysis, human_gate_matrix, containment_and_recovery, and test_plan.
Keep consequential decisions explicit.
These workflows structure analysis and artifacts; they do not authorize autonomous external, personnel, financial, legal, security, or other high-impact action.